On 07/01/2022 12:49, John Keeping wrote:
> On Fri, Jan 07, 2022 at 11:46:45AM +0100, Dietmar Eggemann wrote:
>> On 22/12/2021 20:48, Valentin Schneider wrote:
>>> /*
>>> diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>>> index ef8228d19382..8f3e3a1367b6 100644
>>> --- a/kernel/sched/rt.c
>>> +++ b/kernel/sched/rt.c
>>> @@ -1890,6 +1890,16 @@ static int push_rt_task(struct rq *rq, bool pull)
>>> if (!next_task)
>>> return 0;
>>>
>>> + /*
>>> + * It's possible that the next_task slipped in of higher priority than
>>> + * current, or current has *just* changed priority. If that's the case
>>> + * just reschedule current.
>>> + */
>>> + if (unlikely(next_task->prio < rq->curr->prio)) {
>>> + resched_curr(rq);
>>> + return 0;
>>> + }
>>
>> IMHO, that's the bit which prevents the BUG.
>>
>> But this would also prevent the case in which rq->curr is an RT task
>> with lower prio than next_task.
>>
>> Also `rq->curr = migration/X` goes still though which is somehow fine
>> since find_lowest_rq() bails out for if (task->nr_cpus_allowed == 1).
>>
>> And DL tasks (like sugov:X go through and they can have
>> task->nr_cpus_allowed > 1 (arm64 slow-switching boards with shared
>> freuency domains with schedutil). cpupri_find_fitness()->convert_prio()
>> can handle task_pri, p->prio = -1 (CPUPRI_INVALID) although its somehow
>> by coincidence.
>>
>> So maybe something like this:
>
> Do you mean to replace just the one hunk from Valentin's patch with the
> change below (keeping the rest), or are you saying that only the change
> below is needed?
The latter.
I think Valentin wanted to see if something like this can also occur via
sched_setscheduler() and maybe for this changes in switched_from_[rt/dl]
will be necessary.
>> @ -1898,6 +1898,11 @@ static int push_rt_task(struct rq *rq, bool pull)
>> if (!pull || rq->push_busy)
>> return 0;
>>
>> + if (rq->curr->sched_class != &rt_sched_class) {
>> + resched_curr(rq);
>> + return 0;
>> + }
>> +
>> cpu = find_lowest_rq(rq->curr);
>>
>> [...
